Ontology-Constrained Neural Reasoning in Enterprise Agentic Systems: A Neurosymbolic Architecture for Domain-Grounded AI Agents

Thanh Luong Tuan, Abhijit Sanyal · 2026

Read Ontology-Constrained Neural Reasoning in Enterprise Agentic Systems: A Neurosymbolic Architecture for Domain-Grounded AI Agents by Thanh Luong Tuan, Abhijit Sanyal (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-00555.txt.

Introduction

Thanh and Sanyal present a neurosymbolic design for enterprise agents that keeps reasoning inside domain rules. Enterprise use of language models is held back by hallucination, domain drift, and no way to enforce regulatory compliance at the reasoning step. Their design runs on the Foundation AgenticOS platform and uses a three-layer ontology framework of Role, Domain, and Interaction ontologies to ground agents. Current systems constrain agent inputs but not outputs, so the authors add output-side checks: response checking, reasoning verification, and compliance enforcement.

Catalog record

CREATORThanh Luong Tuan, Abhijit Sanyal
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE55064
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.00555

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Ontology-Constrained Neural Reasoning in Enterprise Agentic Systems: A Neurosymbolic Architecture for Domain-Grounded AI Agents Thanh Luong Tuan, Abhijit Sanyal Abstract Enterprise adoption of Large Language Models (LLMs) is constrained by hallucination, domain drift, and the inability to enforce regulatory compliance at the reasoning level. We present a neurosymbolic architecture implemented within the Foundation AgenticOS ( FAOS ) platform that addresses these limitations through ontology-constrained neural reasoning . We introduce a three-layer ontological framework—Role, Domain, and Interaction ontologies—grounding LLM-based enterprise agents. We formalize asymmetric neurosymbolic coupling : current enterprise systems constrain agent inputs (context assembly, tool discovery, governance thresholds) but not outputs, and we propose mechanisms extending this coupling to output-side validation (response checking, reasoning verification, compliance enforcement). A controlled experiment (1,800 runs across five industries and three LLMs: Claude Sonnet 4, Qwen 2.5 72B, Gemma 4 26B) finds ontology-coupled agents significantly outperform ungrounded agents on Metric Accuracy ( p < .001 p<.001 ) and Role Consistency ( p < .001 p<.001 ) across all three models with large effect sizes (Kendall’s W = .46 W=.46 – .64 .64 ). Improvements are greatest where LLM parametric knowledge is weakest—particularly in Vietnam-localized domains, where ontology lift is 2 × 2\times that of English domains. Contributions: (1) a formal three-layer enterprise ontology model; (2) a taxonomy of neurosymbolic coupling patterns; (3) ontology-constrained tool discovery via SQL-pushdown scoring; (4) a proposed framework for output-side ontological validation; (5) empirical evidence for the inverse parametric knowledge effect—ontological grounding value is inversely proportional to LLM training-data coverage of the domain; (6) cross-model replication establishing model-independence; (7) a production system serving 22 industry verticals with 650+ agents. Keywords: Neurosymbolic AI, Enterprise Ontology, Large Language Models, Agentic AI, Domain-Driven Design, Knowledge-Grounded Reasoning, Multi-Agent Systems 1 Introduction The rapid proliferation of Large Language Models (LLMs) in enterprise settings has exposed a fundamental tension: LLMs excel at natural language understanding and generation but lack the formal semantic grounding required for reliable operation in regulated industries [ 32 , 31 ] . An insurance underwriter that hallucinates policy terms, a financial advisor that confuses Basel III ratios, or a healthcare agent that misapplies HIPAA guidelines poses a compliance and liability risk, not just a quality problem. The “neuro-symbolic debate” has persisted for decades [ 23 , 29 ] : neural approaches offer flexibility and generalization; symbolic approaches provide verifiability and formal guarantees. What is new is the enterprise context —agents must operate within complex organizational structures, adhere to industry-specific regulations, and produce outputs that withstand audit scrutiny. We contend the resolution lies in principled integration: a neurosymbolic architecture where formal ontologies constrain, guide, and verify neural reasoning. This paper presents such an architecture, implemented within the Foundation AgenticOS ( FAOS ) platform, and proposes extensions toward advancing neurosymbolic enterprise AI. 1.1 Enterprise Grounding Enterprise AI agents face a unique variant of the grounding problem [ 28 ] . Unlike general-purpose assistants, enterprise agents must: (i) Speak the domain language : Use industry-specific terminology correctly (e.g., “combined ratio” in insurance, “net revenue retention” in SaaS, “average length of stay” in healthcare). (ii) Reason within regulatory bounds : Ensure outputs comply with applicable regulations (SOX, Basel III/IV, HIPAA, GDPR, EU AI Act [ 16 ] ). (iii) Follow organizational workflows : Respect handoff protocols, approval chains, and escalation paths defined by the organization. (iv) Adapt to role-specific perspectives : A CFO and a product manager examining the same data should receive analyses framed through their respective decision-making lenses. Current approaches address these requirements through prompt engineering [ 58 , 56 ] , retrieval-augmented generation (RAG) [ 35 , 22 ] , and fine-tuning [ 30 ] . While effective for individual dimensions, none provides a unified formal framework for enterprise grounding. Prompt engineering is brittle and unverifiable; RAG retrieves but does not reason; fine-tuning is expensive and domain-locked. 1.2 Contributions This paper contributes: (1) a formal three-layer enterprise ontology model \mathcal{O}=\langle\mathcal{R},\mathcal{D},\mathcal{I}\rangle capturing role-specific reasoning, domain concepts, and interaction protocols ( Section 3 ); (2) a taxonomy of neurosymbolic coupling patterns distinguishing input-side, process-side, and output-side coupling and characterizing current practice as predominantly input-side ( Section 4 ); (3) ontology-constrained tool discovery via SQL-pushdown scoring over domain hierarchies, achieving sub- 100 ms discovery across 600 + registered skills ( Section 6 ); (4) a proposed closed-loop framework for output-side ontological validation ( Section 7 ); (5) a 1,800-run controlled experiment across five regulated industries (including two Vietnamese-language domains) and three LLM architectures, demonstrating significant improvements on Metric Accuracy and Role Consistency ( p<.001 ) and identifying an inverse parametric knowledge effect where grounding value is greatest where LLM training coverage is weakest ( Section 8 ); (6) cross-model replication on Claude Sonnet 4, Qwen 2.5 72B, and Gemma 4 26B establishing model-independence ( Section 9.4 ); and (7) production evidence from a system implementing input-side coupling across 22 industry verticals, 650+ agents, and 7 bounded contexts ( Section 5 ). 2 Background and Related Work 2.1 Neurosymbolic AI Neural–symbolic integration has been a persistent research theme since the connectionist-symbolic debates of the 1980s [ 19 ] , catalyzed recently by deep learning maturation, knowledge representation standards (OWL, RDF, SKOS), and the emergence of LLMs as general-purpose reasoning engines. 23 distinguish symbolic[neural] , neural[symbolic] , and neural:symbolic (bidirectional) integration; our work falls primarily in neural[symbolic] with proposals toward neural:symbolic . A systematic review of 167 NeSyAI papers [ 11 ] confirms the convergence. 29 argue that ontologies provide a natural formalism for this integration—type hierarchies, property constraints, and inference rules that can bound neural computation; we extend the argument to the enterprise domain, where ontologies must also capture organizational roles and interaction protocols. 2.2 Knowledge-Grounded Language Models Knowledge grounding for LLMs has been pursued through four main mechanisms. Retrieval-Augmented Generation (RAG) augments LLM context with retrieved documents [ 35 , 22 ] , addressing factual grounding but not structural or regulatory constraints. Knowledge Graph Integration [ 47 , 64 ] provides relational grounding but typically lacks the formal semantics of ontological reasoning. Ontology-Enhanced Prompting uses ontological definitions to structure prompts [ 9 ] ; 36 extend this with OntoTune (WWW 2025), aligning LLMs through in-context learning on SNOMED CT. Constrained Decoding restricts output tokens to grammatical or structural constraints [ 60 ] —syntactic rather than semantic control. Broader surveys of tool-augmented LLMs [ 46 ] and hybrid architectures [ 45 ] argue for neural–symbolic combination; 33 demonstrate that LLMs struggle with planning without external symbolic scaffolding. Empirical work has begun to validate these approaches quantitatively: 2 survey KG-augmentation for hallucination reduction; 25 show ontology-grounded LLMs exhibit reduced digression in industrial conversation; 8 find ontology-guided KG construction outperforms vector retrieval on specialized domains; and 50 propose a KG-based framework for triple-level hallucination measurement. Our work synthesizes these within a unified multi-layer ontological framework and targets enterprise-specific tasks that test the boundaries of parametric knowledge. 2.3 Ontology-Grounded RAG and Hallucination Mitigation Recent work demonstrates dramatic improvements from ontological grounding over baseline RAG: 52 propose OG-RAG (EMNLP 2025), achieving +55\% fact recall and +40\% response correctness across four LLMs via ontology-anchored hypergraph retrieval; 3 report 98\% clinical QA accuracy (vs. 37\% for ChatGPT-4) through RDF/OWL knowledge graphs, reducing hallucination from 63\% to 1.7\% ; 4 ground each reasoning step

Read the complete text